PaperlayerView exact revision on arXiv

Source-linked AI summary

Think in Latent, Explain in Language: Self-Explainable Latent Reasoning

Dayuan Zhao, Shengcao Cao, Yu-Xiong Wang, Liang-Yan Gui

arXiv:2608.13570v1cs.CLcs.AIcs.LG

TL;DR

Latent reasoning is efficient but difficult to supervise and interpret. SELR trains one model with answer and CoT losses to reason in latent space and decode its own thoughts, achieving gains in accuracy and token efficiency across LLM and VLM tasks.

  • Problem

    Latent reasoning lacks generally agreed supervision and direct interpretability because its continuous thoughts have no human-readable ground truth.

  • Method

    SELR jointly trains a single model with Answer and CoT losses to produce accurate latent reasoning and translate its representations into human-readable steps.

  • Results

    Across LLM and VLM tasks, SELR improves reasoning accuracy and token efficiency while decoding its own continuous thoughts.

  • Takeaways & Limitations

    SELR provides a practical path toward latent reasoning that combines computational efficiency with self-contained explainability.

  • Takeaways & Limitations

    SELR requires a predefined inference-time latent-token budget because autonomously learning when to stop remains difficult.

Abstract

from arXiv · show

Latent reasoning has emerged as a powerful alternative to text-based Chain-of-Thought (CoT), offering significant gains in computational efficiency by compressing verbose reasoning into compact embeddings. However, compressing reasoning into the latent space renders the thinking opaque, hindering its interpretability. Current methods present a stark trade-off: they either function as unexplainable ''black boxes'' (e.g., Coconut), where the latent reasoning is not human-readable, or rely on separate post-hoc decoders for explainability (e.g., Heima), introducing architectural overhead and decoupling the explanation from the actual reasoning process. In this work, we present a unified framework for Self-Explainable Latent Reasoning (SELR) that trains a single model to perform efficient and inherently explainable latent reasoning. Our core contribution is a novel multi-task training objective that optimizes for two goals simultaneously: (1) an Answer Loss that optimizes the latent reasoning trajectory to produce accurate final answers, and (2) a CoT Loss that explicitly trains the same model to decode its own latent representations back into human-understandable reasoning steps. This design ensures that generated latent representations are both task-effective and semantically interpretable, eliminating the need for external decoders. We validate the effectiveness of SELR on both Large Language Models (LLMs) and Vision-Language Models (VLMs), demonstrating that SELR achieves superior token efficiency and accuracy compared to baselines, while uniquely providing self-contained explainability without auxiliary models. Project page is available at https://jasondayuan.github.io/SELR/.

1. Introduction

Latent reasoning can improve efficiency over verbose text-based Chain-of-Thought, but lacks human-interpretable supervision and direct interpretability. SELR addresses both problems by training one model to reason in latent space and decode its own latent thoughts, with experiments extending the approach from LLMs to VLMs.

  • Motivation: Text-based Chain-of-Thought can be excessively verbose and constrained by a fixed, discrete vocabulary, motivating reasoning in latent space.
  • Challenges: Latent reasoning lacks human-interpretable supervision because continuous thoughts have no textual ground truth, and its internal process is difficult to understand directly.
  • Challenges: Token-level probing is poorly human-readable, while separate post-hoc decoders add parameters and may decouple explanations from the reasoning process.
  • SELR Framework: SELR trains a single model with Answer Loss and CoT Loss to perform efficient latent reasoning and translate its own latent thoughts into human-readable reasoning.
  • SELR Framework: The CoT Loss supplies text-based supervision and encourages latent representations aligned with human logic, making thoughts both task-effective and inherently explainable.
  • Experiments: Experiments identify training strategies on lightweight LLMs and show their generalization to VLMs, including simultaneous reasoning-accuracy and token-efficiency improvements for Qwen2.5-VL.

2. Related Work

Related work spans explicit Chain-of-Thought reasoning, which elicits step-by-step solutions through prompting and decoding strategies, and latent-space reasoning, which compresses reasoning into continuous hidden states for greater computational efficiency.

  • Chain-of-Thought Reasoning: CoT elicits step-by-step reasoning through few-shot examples or simple prompts in large language models.The framework is enhanced by self-consistency decoding and generalized through search-based structures such as Tree of Thoughts.
  • Chain-of-Thought Reasoning: CoT has also been adapted for vision-language tasks.
  • Latent Space Reasoning: Coconut moves reasoning into a continuous latent space by feeding the last hidden state back as a continuous thought.This approach addresses the computational inefficiency of explicit CoT and enables advanced, non-deterministic reasoning such as latent Breadth-First Search.

3. Approach

SELR trains latent reasoning to support accurate answers while remaining self-explainable through a unified multi-task framework. It generates latent thoughts by feeding hidden states back as embeddings, then applies answer and CoT supervision under fixed or variable reasoning budgets.

  • Answer Loss: The Answer Loss uses cross-entropy over ground-truth answer tokens conditioned on the image, question, and latent thoughts.This loss supervises the model to arrive at the correct answer through latent-space reasoning.
  • Latent Space Reasoning: SELR generates latent thoughts autoregressively by appending each final hidden state directly to the input embedding sequence, bypassing the language-model head.The process initializes embeddings from image features, the question, and <bot>, then stores the appended hidden states as Elatent.
  • Thinking Budget: SELR supports fixed-length and variable-length latent reasoning, but variable-length models still require a predefined inference budget because autonomous termination is difficult to train.Fixed-length models use a constant llatent, whereas variable-length training matches llatent to the instance’s K reasoning steps.
  • CoT Loss: The CoT Loss decodes latent reasoning into human-readable text through either Full CoT Loss over the entire sequence or Single Step Loss over a corresponding step.Single Step Loss requires variable latent reasoning length because it relies on a one-to-one correspondence between latent thoughts and textual reasoning steps.
  • CoT Loss: During CoT decoding, the model receives only latent thoughts, forcing them to form a self-contained information bottleneck that carries necessary solution context.The original image and question are withheld during decoding.

4. Experiments

Experiments show that SELR improves accuracy and token efficiency across LLM and VLM settings while providing faithful self-explanations through a unified model. Ablations attribute these gains to CoT supervision, multi-stage training, and fixed latent lengths, while identifying limitations on tasks requiring extended generation.

  • Experiment Setup: SELR trains VLMs on LLaVA-CoT-100k’s 100k image-text pairs with summary, caption, and reasoning stages, and LLMs on GSM8k-Aug’s 385k generated math problems.These datasets support experiments across vision-language and language-only reasoning tasks.
  • Main Results: 0.86% is SELR’s best average-performance gain over the original model, whereas Heima shows a 1.92% drop.The result indicates that SELR preserves or improves general and reasoning capabilities despite changing the generation paradigm.
  • Main Results: Approximately 13 tokens is the average response length of SELR multi-stage variants, representing over 70% reduction from the base model and outperforming Heima’s 22.84-token average.SELR improves efficiency while maintaining the performance gains reported in the experiments.
  • LLM Results: SELR (Multi) outperforms Coconut and the single-stage variant on in-domain and out-of-domain datasets, surpasses CoLaR on GSM8k and GSM-Hard, and uses fewer reasoning tokens.The multi-stage curriculum remains competitive with CoLaR while achieving shorter reasoning length.
  • LLM Results: 42.46% is SELR’s GSM8k accuracy with 6 latent tokens, compared with 10.31% for CoT-SFT constrained to the same budget.CoT-SFT normally generates over 150 text tokens per question, so the gap reflects reasoning budget rather than reasoning quality.
  • VLM Explainability: SELR’s decoded summaries, captions, and reasoning are more similar to ground truth than Heima, with step-wise decoding outperforming full-sequence decoding for VLMs.Heima uses separate decoders for its latent thoughts, whereas SELR unifies reasoning and self-explanation in one model.
  • Faithfulness: Correct predictions achieve approximately 0.5 decoded-CoT faithfulness versus approximately 0.26 for incorrect predictions, with SELR (Multi, Exponential) highest among variants.This correlation indicates that decoded CoT reflects meaningful reasoning content rather than generic post-hoc outputs.
  • Faithfulness: 0.2097 is SELR (Single)’s MathVista consistency score versus Heima’s 0.1671, a +25.5% relative improvement enabled by unified reasoning and decoding parameters.The result, together with decoding-quality comparisons, supports stronger faithfulness than Heima.

5. Conclusion · Appendix · A. Implementation Details

The conclusion presents SELR as a unified model that combines answer and CoT losses to achieve efficient, self-explainable reasoning for VLM and LLM tasks. The appendices describe supplementary experiments, visualizations, alternative configurations, ablations, and implementation details.

  • 5. Conclusion: SELR trains one model to act as both an efficient reasoner and its own translator.Its objective combines a standard answer loss with a CoT loss.
  • 5. Conclusion: The multi-task objective combines answer and CoT losses within a unified training framework.The CoT loss supports the model’s translation of latent reasoning, while the answer loss supports task performance.
  • 5. Conclusion: Experiments report SELR effectiveness on both VLM and LLM tasks.The conclusion states that the method yields performance gains while significantly reducing response length.
  • Appendix: The appendix supplies supplementary material supporting the main paper’s experimental and qualitative analyses.It covers experimental setups, decoded latent-thought examples and visualizations, alternative SFT base configurations, and ablations.
  • Appendix: Appendix C evaluates SELR on alternative SFT base configurations.This evaluation is listed among the appendix’s supplementary investigations.
  • Appendix: Appendices D and E ablate fixed latent lengths and the evaluation protocol for variable-length reasoning.These studies examine methodological choices affecting the framework’s evaluation.
  • A. Implementation Details: The implementation-details section introduces the model training and evaluation details.It frames the section as a description of how the model is trained and assessed.

A.1. Training Configuration · A.2. Inference Settings

Training uses a 15-epoch, two-stage schedule for LLM SELR, while baseline configurations and VLM details are specified separately. Inference uses greedy decoding for LLMs and low-temperature sampling for VLMs, and removing structured XML tags reduces average performance by 5.14%.

  • A.1. Training Configuration: SELR’s LLM training runs for 15 epochs: 10 epochs at learning rate 5e-4 followed by 5 epochs at 1e-4.The schedule is divided into two distinct stages.
  • A.1. Training Configuration: Coconut is reimplemented with a matched 15-epoch budget, using 3 initial epochs and 3 epochs for each of three intermediate and one final stage.The reproduction follows the original curriculum setup and official repository code.
  • A.1. Training Configuration: SELR training requires about 30 hours, compared with over 50 hours for Coconut on four NVIDIA A100 GPUs.The reported comparison concerns training cost under the described configurations.
  • A.1. Training Configuration: Finetuning Qwen2.5-VL-3B-Instruct on LLaVA-CoT-100k with structured XML tags preserves the recommended training format.The tags include markers such as <SUMMARY> and </SUMMARY>; an untagged variant is also evaluated.
  • A.1. Training Configuration: Removing structured XML tags causes a 5.14% drop in average performance for the finetuned Qwen2.5-VL-3B-Instruct model.This result agrees with the LLaVA-CoT paper’s reported degradation without tags.
  • A.2. Inference Settings: LLM inference uses greedy decoding, whereas VLM inference uses sampling with temperature 1e-2, top-p 1e-3, and repetition penalty 1.0.These VLM settings essentially match Qwen2.5-VL-3B-Instruct defaults in VLMEvalKit for fair baseline comparison.

B. Additional Qualitative Results · C. SELR on Qwen2.5-VL-3B-Instruct-SFT · D. Latent Length Ablation

Additional examples show that SELR produces concise, coherent, human-readable reasoning and correct answers, while remaining effective on a weaker Qwen2.5-VL base model. For LLMs, latent length 6 offers the best reasoning-capacity trade-off, whereas longer sequences add little benefit.

  • B. Additional Qualitative Results: SELR gives concise responses and improves token efficiency over the base model in additional qualitative examples.These examples are presented in Figure 5.
  • B. Additional Qualitative Results: SELR produces human-readable, coherent reasoning that leads to correct answers in further examples.These examples are presented in Figure 6.
  • C. SELR on Qwen2.5-VL-3B-Instruct-SFT: SELR variants are compared under identical optimization settings when fine-tuned on Qwen2.5-VL-3B-Instruct-SFT.The comparison uses the SFT baseline and SELR variants; Heima follows its proposed curriculum.
  • C. SELR on Qwen2.5-VL-3B-Instruct-SFT: Table 11 evaluates SELR (Single), SELR (Multi, Uniform), and SELR (Multi, Exponential) using Qwen2.5-VL-3B-Instruct-SFT (w/o Structured Tags).The training hyperparameters are identical to those in Table 10, isolating latent-reasoning gains from target-dataset fine-tuning.
  • C. SELR on Qwen2.5-VL-3B-Instruct-SFT: Even with Qwen2.5-VL-3B-Instruct-SFT (w/o Structured Tags), SELR models surpass Qwen2.5-VL-3B-Instruct-SFT.The weaker base model is used to avoid shifting the generation pattern caused by adding or removing tags.
  • C. SELR on Qwen2.5-VL-3B-Instruct-SFT: A decoded CoT can be correct while the direct answer is wrong, showing that the decoded CoT may not represent the model’s true reasoning path.Figure 4 illustrates this unfaithful CoT decoding case.
  • D. Latent Length Ablation: 6 latent thoughts provide the optimal balance for LLMs; length 2 or 4 limits reasoning capacity, while length 8 yields no significant improvement.The ablation reduces training to 2 epochs for efficiency and otherwise follows the SELR (Single) configuration in Table 9; VLM ablation is not viable because Single Step Loss requires one-to-one latent-to-reasoning-step correspondence.

E. Evaluating Variable Latent Length Methods · F. Prompt for GPT-4o Evaluation

The variable-latent-length evaluation uses fixed inference budgets and reports the best accuracy across tested budgets, yet generally trails the fixed-length strategy. GPT-4o evaluates decoded-CoT similarity using Heima’s identical prompt for fair comparison.

  • E. Evaluating Variable Latent Length Methods: E. Evaluating Variable Latent Length Methods: An inference budget b generates exactly b latent thought tokens after <bot> before termination.Because predicting <eot> is difficult, evaluation cuts off latent reasoning with a fixed budget.
  • E. Evaluating Variable Latent Length Methods: E. Evaluating Variable Latent Length Methods: Tables 7 and 8 report each benchmark’s optimal accuracy across tested budgets as an upper-bound estimate.The reported result assumes oracle selection of the best budget.
  • E. Evaluating Variable Latent Length Methods: E. Evaluating Variable Latent Length Methods: Even with oracle budget selection, variable latent length generally underperforms the proposed fixed latent length strategy.This comparison concerns the optimal accuracy across tested budgets.
  • E. Evaluating Variable Latent Length Methods: E. Evaluating Variable Latent Length Methods: The SELR (Single) latent-length ablation trains for 2 epochs across varying lengths, with l_latent = 6 best on most datasets.This finding is reported for LLM benchmarks.
  • E. Evaluating Variable Latent Length Methods: E. Evaluating Variable Latent Length Methods: Table 13 evaluates single-stage SELR trained with variable latent length using fixed budgets from 1 to 5 latent tokens.The optimal result for each dataset is bolded.
  • F. Prompt for GPT-4o Evaluation: F. Prompt for GPT-4o Evaluation: GPT-4o assesses similarity between decoded CoTs and ground-truth CoTs using the identical prompt employed by Heima.The prompt is reproduced in Algorithm 2 for clarity and fair comparison.

G. Inference Latency

On MathVista, SELR (Single) converts token reduction into lower wall-clock inference latency while maintaining the highest reported accuracy. It reaches 65.10% accuracy at 0.81s per sample, an 8.7× speedup over SFT.

  • 65.10% accuracy is achieved by SELR (Single), the highest reported result on MathVista.
  • 0.81s per sample is SELR (Single)’s lowest reported inference latency on MathVista.
  • 8.7× speedup over the SFT baseline is achieved by SELR (Single).

H. Qwen2.5-VL-7B-Instruct Experiments

Experiments on Qwen2.5-VL-7B-Instruct show that SELR generalizes across model scales, improving accuracy while substantially reducing token usage.

  • Qwen2.5-VL-7B-Instruct Experiments: SELR improved average accuracy over the original Qwen2.5-VL-7B-Instruct model by +0.11 and reduced token usage by over 80%.SFT degraded accuracy despite generating far more tokens.

I. Controlled-Budget LLM Comparison · J. Generalization to VisualWebInstruct-Verified · K. Discussion on Reasoning Step Ordering

SELR’s advantage over textual CoT persists under equal reasoning budgets and extends to a differently structured visual-reasoning dataset. In VLMs, its fixed summary–caption–reasoning order supports natural progression and positional specialization, while LLM reasoning remains unstructured.

  • I. Controlled-Budget LLM Comparison: CoT-SFT generates over 150 reasoning tokens on GSM8k, whereas SELR uses 6 latent tokens; controlled comparisons therefore truncate CoT-SFT to 6 tokens plus the answer and report zero-token performance.The comparison tests whether performance differences reflect reasoning mechanisms or merely reasoning budgets.
  • J. Generalization to VisualWebInstruct-Verified: SELR outperforms CoT-SFT on 5 of 6 VisualWebInstruct-Verified benchmarks, demonstrating compatibility with both structured and unstructured reasoning annotations.The result is presented as evidence that the improvements are not an artifact of a particular dataset structure.
  • I. Controlled-Budget LLM Comparison: 42.46% on GSM8k is achieved by SELR versus 10.31% by equally budgeted CoT-SFT, showing more efficient reasoning encoding.The passage describes this as a 4× advantage for SELR and attributes CoT-SFT’s usual advantage to its much longer reasoning trace.
  • J. Generalization to VisualWebInstruct-Verified: SELR and CoT-SFT are evaluated on VisualWebInstruct-Verified using Qwen2.5-VL-3B-Instruct to test generalization beyond LLaVA-CoT-100k’s source and structured format.VisualWebInstruct-Verified is described as a more recent VLM reasoning dataset with a completely different source and format.
  • K. Discussion on Reasoning Step Ordering: In VLMs, SELR fixes the reasoning-step order as summary →caption →reasoning, following LLaVA-CoT-100k and prior methods including LLaVA-CoT and Heima.This ordering is inherited from LLaVA-CoT-100k and is also adopted by LLaVA-CoT and Heima.
  • K. Discussion on Reasoning Step Ordering: The fixed VLM order provides a natural visual-question-answering flow and lets each latent position specialize, whereas randomization would increase learning difficulty.The sequence moves from understanding the question to describing visual content and then reasoning toward the answer.
  • K. Discussion on Reasoning Step Ordering: The fixed ordering applies only to VLMs; in LLMs, SELR uses unstructured CoT steps that follow the natural reasoning sequence without predefined structure.Using the same VLM ordering as prior LLaVA-CoT-100k methods also ensures fair comparison.

L. Limitations

SELR makes latent reasoning decodable but does not guarantee faithful transparency, and its fixed reasoning length limits flexibility. Future work should improve faithfulness, explainability, and adaptive stopping.

  • Faithfulness and transparency: SELR enables latent reasoning to be decoded, unlike prior latent methods, but faithfulness remains an open challenge.Text-CoT models can be systematically unfaithful, and SELR’s decoded reasoning may not reflect the model’s true internal process.
  • Faithfulness and transparency: “Self-Explainable” denotes self-decodable latents with meaningful faithfulness signals, not guaranteed full reasoning transparency.Decoded CoT can occasionally be correct even when the model’s direct answer is incorrect, indicating potential misalignment.
  • Reasoning length and stopping: SELR uses a fixed latent reasoning length, which may be suboptimal, while predicting the <eot> token remains difficult for the LM head.A learnable stopping mechanism could let the model determine when latent thinking should end, but this is outside the current scope.
  • Future work: Future work aims to improve latent reasoning’s explainability and flexibility while retaining its efficiency and efficacy.The paper identifies more advanced, learnable stopping as one direction for addressing current limitations.
Loading 2608.13570v1…